Seasonal Naive Forecasts

Why Use a Seasonal Naive Forecast?

When a time series has a recurring pattern, a forecast can use the most recent observation from the same seasonal position. For example, a monthly forecast for next January can use the latest observed January value, while a quarterly forecast for next fourth quarter can use the latest observed fourth-quarter value. This is the seasonal naive method: a simple benchmark that repeats the last observed seasonal cycle.

The method makes no adjustment for a trend or a changing seasonal pattern. It is useful when the most recent cycle is a reasonable guide to the next one, and as a baseline against which more elaborate forecasts can be compared. The forecasts are determined by the observed values and the seasonal period; no probability distribution or parameter estimation is required.

Definition: Let $y_1,\ldots,y_T$ be the observed time series through forecast origin $T$, and let $m$ be the number of time periods in one seasonal cycle. For an integer horizon $h\geq 1$, the seasonal naive forecast is the latest observed value in the same seasonal position as the target period: $$ \hat{y}_{T+h\mid T}=y_{T+h-m\lceil h/m\rceil}. $$ The formula applies when its observation index is at least 1. In particular, if at least one complete seasonal cycle has been observed ($T\geq m$), it is defined for every positive forecast horizon.

Finding the Matching Seasonal Position

The seasonal period $m$ is the number of observations between equivalent seasonal positions: typically $m=12$ for monthly data with an annual cycle, or $m=4$ for quarterly data with an annual cycle. The forecast horizon $h$ counts periods forward from the last observation. Thus, $h=1$ means the next period, and $h=13$ in monthly data means thirteen months ahead.

The ceiling in the formula counts how many seasonal cycles must be stepped back from the target period to reach an observation no later than $T$. For horizons 1 through $m$, it is one cycle back. For horizons $m+1$ through $2m$, it is two cycles back. Consequently, forecasts beyond the first cycle reuse the same observed values rather than recursively using forecasts.

Identify the seasonal period.
Set $m$ to the number of periods in a complete cycle, such as 12 for monthly annual seasonality or 4 for quarterly annual seasonality.

Specify the forecast origin and horizon.
Use $T$ for the index of the last observed value and $h$ for the number of periods ahead of it.

Locate the past seasonal match.
Calculate $k=\lceil h/m\rceil$, then find the observation with index $T+h-km$.

Check that the observation exists.
The index must be at least 1 and no greater than $T$. If it is less than 1, the available history does not contain that seasonal position.

Why the Index Formula Works

The target is at time $T+h$. Observations in the same seasonal position occur every $m$ periods, so candidates for the matching past observation have indices $T+h-mk$, where $k$ is a positive integer number of cycles. We need the latest such candidate that is observed, so its index must be no greater than $T$. The condition

T+h-mk\leq T \quad\Longleftrightarrow\quad k\geq \frac{h}{m}

The smallest positive integer satisfying this inequality is $k=\lceil h/m\rceil$. Using this smallest $k$ gives the latest matching seasonal observation, hence the index $T+h-m\lceil h/m\rceil$. If $T\geq m$, this index is valid for all $h\geq1$: for $h$ within any block of $m$ horizons, the smallest possible index is $T+1-m$, which is at least 1. The index is also at most $T$ by the choice of $k$.

Worked Example: Monthly Forecasts Across Two Cycles

A hypothetical home-services provider records monthly service requests. There are $T=15$ observations and the seasonal period is $m=12$ months. The observations, in order, are $82, 86, 91, 104, 110, 124, 130, 128, 115, 102, 94, 89, 88, 92, 97$. Find forecasts at horizons $h=1,6,9,12,13,$ and $15$. This is a deterministic forecasting rule, so no probability distribution is specified or needed.

Method. For each horizon, calculate $k=\lceil h/12\rceil$, then use the observation at index $15+h-12k$.

At $h=1$, $k=\lceil1/12\rceil=1$, so the index is $15+1-12=4$. Therefore $\hat y_{16\mid15}=y_4=104$.

At $h=6$, $k=\lceil6/12\rceil=1$, so the index is $15+6-12=9$. Therefore $\hat y_{21\mid15}=y_9=115$.

At $h=9$, $k=\lceil9/12\rceil=1$, so the index is $15+9-12=12$. Therefore $\hat y_{24\mid15}=y_{12}=89$.

At $h=12$, $k=\lceil12/12\rceil=1$, so the index is $15+12-12=15$. Therefore $\hat y_{27\mid15}=y_{15}=97$.

At $h=13$, $k=\lceil13/12\rceil=2$, so the index is $15+13-24=4$. Therefore $\hat y_{28\mid15}=y_4=104$.

At $h=15$, $k=\lceil15/12\rceil=2$, so the index is $15+15-24=6$. Therefore $\hat y_{30\mid15}=y_6=124$.

Check. The first forecast uses the most recently observed value from the same month, $y_4$; twelve months later, the forecast again uses $y_4$. Likewise, $h=12$ uses the final observed value, since the target is one annual cycle after it. A common mistake is to use $y_{T+1-m}=y_4$ for every horizon in the first year. That is only the match for $h=1$; at $h=6$, for example, the correct matching index is 9, not 4.

Worked Example: Quarterly Forecasts at Multiple Horizons

A hypothetical insurer tracks quarterly counts of roadside-assistance claims. The ten observations are $40,45,55,50,44,49,62,56,47,52$, so $T=10$. Use $m=4$ and calculate forecasts at horizons $h=1,2,4,5,$ and $8$. As with any seasonal naive calculation, no probability distribution is required.

Method. The matching index is $10+h-4\lceil h/4\rceil$.

At $h=1$, $k=\lceil1/4\rceil=1$, giving index $10+1-4=7$. Thus $\hat y_{11\mid10}=y_7=62$.

At $h=2$, $k=\lceil2/4\rceil=1$, giving index $10+2-4=8$. Thus $\hat y_{12\mid10}=y_8=56$.

At $h=4$, $k=\lceil4/4\rceil=1$, giving index $10+4-4=10$. Thus $\hat y_{14\mid10}=y_{10}=52$.

At $h=5$, $k=\lceil5/4\rceil=2$, giving index $10+5-8=7$. Thus $\hat y_{15\mid10}=y_7=62$.

At $h=8$, $k=\lceil8/4\rceil=2$, giving index $10+8-8=10$. Thus $\hat y_{18\mid10}=y_{10}=52$.

Check. Since $y_7$ is a third-quarter observation, it supplies both the one-quarter-ahead forecast and the five-quarter-ahead forecast. The eight-quarter-ahead forecast uses $y_{10}$, the latest observed value in the same quarterly position as the target. The key trap is confusing a horizon of four with a four-step-back observation: at $h=4$, the forecast is $y_{10}$, not $y_6$.

When the Available History Is Too Short

The formula cannot create an observation from a seasonal position that has never been observed. A complete cycle, $T\geq m$, is a sufficient condition for forecasts at every positive horizon. If fewer than $m$ observations are available, some forecast horizons may still have a matching observation, but others will not. For each requested horizon, calculate the index and verify it is at least 1 before assigning a forecast.

For example, with $T=3$, $m=4$, and $h=1$, the index is $3+1-4\lceil1/4\rceil=0$. There is no observed $y_0$, so the seasonal naive forecast is undefined. At $h=2$, the index is $3+2-4=1$, which is a valid match. The method therefore supports that horizon, but not every horizon from this incomplete history.

Worked Example: A Missing Seasonal Position

A hypothetical monitoring system has collected only three quarterly readings: $y_1=18$, $y_2=23$, and $y_3=21$. The seasonal period is $m=4$, but the fourth quarter has not yet been observed. Determine whether seasonal naive forecasts can be made for $h=1,2,3,$ and $4$. No probability model is assumed.

Method. For each horizon, use the index $3+h-4\lceil h/4\rceil$, and reject any index below 1.

For $h=1$, $k=\lceil1/4\rceil=1$, and the index is $3+1-4=0$. Thus, no seasonal naive forecast is available: $y_0$ is not observed.

For $h=2$, $k=1$, and the index is $3+2-4=1$. The forecast is $\hat y_{5\mid3}=y_1=18$.

For $h=3$, $k=1$, and the index is $3+3-4=2$. The forecast is $\hat y_{6\mid3}=y_2=23$.

For $h=4$, $k=1$, and the index is $3+4-4=3$. The forecast is $\hat y_{7\mid3}=y_3=21$.

Check. The missing value at $h=1$ corresponds to the seasonal position represented by $y_4$, which is not yet in the data. A common mistake is to apply the formula without checking the index and report $y_0$ as a forecast. The formula requires an observed value: the correct conclusion at $h=1$ is that this method cannot provide a forecast from the available history.

Interpretation and Common Traps

A seasonal naive forecast repeats the latest observed seasonal cycle, not necessarily the entire history and not necessarily the last observation at every horizon. For $1\leq h\leq m$, it takes one of the last $m$ observations. At longer horizons it cycles back to those same observations. This means that, after the first $m$ forecasts, the sequence of point forecasts repeats with period $m$.

This behavior is useful when the seasonal pattern is persistent and the most recent cycle is representative. It can be unsuitable when the series has a substantial trend, when seasonal effects are changing, or when an unusual event makes the last cycle unrepresentative. The method does not estimate a trend or smooth noisy values. Its simplicity is also an advantage: it provides a clear reference forecast for judging whether a more complex approach adds useful predictive performance.

Confusing the horizon with the seasonal period. A horizon of $m$ uses the last observed value $y_T$; it does not mean stepping back $m$ observations from $T$.
Reusing the same value for every horizon. Forecasts at different seasonal positions generally use different observations from the last cycle.
Using forecasts as if they were observed values. The seasonal naive rule refers back to the observed series, even for horizons beyond $m$.
Ignoring the history requirement. An index below 1 means no matching observation is available; it is not a valid forecast.

Key takeaway: For a forecast $h$ periods beyond time $T$, step back $\lceil h/m\rceil$ complete seasonal cycles from the target time. Use the resulting observed value only if its index is at least 1. If $T\geq m$, the rule is defined at every positive horizon.

Check Your Understanding

For each problem, identify the matching observation and verify that its index is available before giving the forecast.

A quarterly series has $T=8$, $m=4$, and observations $12,15,18,20,14,17,21,23$. Find the seasonal naive forecasts at $h=1,3,4,$ and $5$. (Answer: $14,21,23,14$, respectively.)
A monthly series has $T=14$, $m=12$, and observations $31,34,36,39,42,45,47,46,43,40,37,35,33,36$. Find the forecasts at $h=2,12,$ and $13$. (Answer: $y_4=39,\ y_{14}=36,\ y_3=36$, respectively.)
A seasonal series has $T=5$, $m=6$, and observed values $y_1=8,y_2=11,y_3=13,y_4=16,y_5=14$. Is the forecast available at $h=1$? Find it at $h=2$ and $h=6$. (Answer: No forecast at $h=1$, because its index is 0; at $h=2$, the forecast is $y_1=8$; at $h=6$, it is $y_5=14$.)
A quarterly series has $T=11$, $m=4$, and observations $y_1,\ldots,y_{11}=26,28,33,31,29,30,36,34,32,35,39$. Calculate forecasts at $h=6$ and $h=9$. (Answer: At $h=6$, the index is $11+6-8=9$, so the forecast is $y_9=32$; at $h=9$, the index is $11+9-12=8$, so the forecast is $y_8=34$.)
A monthly series has $T=24$, $m=12$, with $y_{13}=52$, $y_{18}=61$, and $y_{24}=49$. Find the forecasts at $h=1,6,12,13,$ and $18$. (Answer: $y_{13}=52,\ y_{18}=61,\ y_{24}=49,\ y_{13}=52,\ y_{18}=61$, respectively.)

What You'll Learn